Limits and opportunities of multispectral data for estimating soil organic carbon (SOC) content in croplands.

Soil Organic Carbon (SOC) can play a role in climate mitigation and is critical to soil functioning, yet assessment of its spatial and temporal variability remains challenging. Advances in soil spectroscopy and increasing availability of multispectral satellite images have raised expectations that bare soil reflectance could support SOC content estimation at high spatial resolution. However, the suitability of such data for either SOC content mapping (spatial variability) and/or SOC monitoring (temporal variability) remains insufficiently understood. The spatial resolution of Sentinel-2 could allow to estimate SOC content at parcel, or intra-parcel spatial support.

Using a unique dataset of 34,418 parcel-level soil analyses, we evaluated whether Sentinel‑2 bare soil composites can support SOC mapping and monitoring in Wallonia (Belgium). The models were used to assess (i) the complementarity of spectral data to information from a set of environmental covariates, (ii) quantify prediction uncertainty, (iii) assess the potential of SOC content estimates for mapping or monitoring at parcel and aggregated spatial levels, and (iv) interpretation of the spectral features selected by the models.

In the context of the Walloon region, environmental and spectral covariates were found to be complementary. The strongest gains in performance were in homogeneous areas where SOC variability is low and poorly captured by spectral data alone. The performance of the model increases when including soil property maps, altitude and agro-ecological zones as covariates. The optimal model estimates SOC content with an RMSE of 2.7 g C kg^-1 at parcel-level and quantile regression methods provided reliable uncertainty estimates. The predictive value of all pairwise combinations of Sentinel-2 bands was evaluated, allowing to select 4 relevant spectral indices. The Minimum Detectable Difference of the model estimates exceeded the expected rates of SOC content change within 10 years of implementing carbon farming practices. This indicates that multispectral bare soil composites are not suitable for SOC content monitoring at parcel-level, as noise may be misinterpreted as temporal SOC content variability.

Interpretation of the selected spectral features revealed that there is a physical explanation for their correlation to SOC content. The selected index NBR2 however, is likely correlated to clay content rather than SOC content, further limiting the use for temporal monitoring. In contrast, our results support the use of bare soil composites for SOC content mapping, providing relevant spatial information, complementary to traditional digital soil mapping covariates. Moreover, the maps allow for estimation of average SOC content at aggregated spatial units, for which uncertainty quantification is essential.

Our results therefore support the use of multispectral bare soil composites for SOC content mapping, while tempering expectations regarding their use for SOC monitoring at parcel-level. Future work should focus on refining mapping approaches and improving understanding of the indirect spectral–soil relationships that underpin SOC predictions as well as the use of hyperspectral imagery for SOC content estimation.